Cowork: context compaction silently drops CONFIRMED work, and the "resume directly" prompt tells the agent to skip re-grounding

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Product note (please read first): this is about Claude Cowork, the desktop-app feature on the Claude platform, NOT the Claude Code CLI. I am filing here because there is no public issue tracker for Cowork and the CLI templates were the closest available. Reporter: paid Max plan, latest desktop app as of 2026-07-17.

What is wrong: When a long Cowork session hits context compaction, the platform replaces the conversation with an auto-generated summary and injects a system instruction telling the agent to "resume directly, do not acknowledge the summary." The summary is lossy and can be inaccurate, and it drops CONFIRMED state:

Work I had already confirmed complete was treated by the resumed agent as still open, forcing me to re-confirm signed-off work.

Worse: a feature the agent had BUILT and I had CONFIRMED days earlier was silently absent from the summary. The resumed agent had no record it existed, built new work on the state that feature depended on, and instructed me to run a live step that undermined the earlier feature. I caught it only because I personally remembered; the agent could not have, because its whole memory of the session was the summary and the summary looked complete.
The summary also cited disposable in-app task-panel numbers as if they were canonical task IDs.

The "resume directly" instruction works directly against any user-side discipline to re-verify state after a truncation. I built an explicit re-onboarding routine and the platform prompt tells the agent to skip exactly that step.

What Should Happen?

Expected behavior: Compaction preserves confirmed features and decisions verbatim, not just the most recent work.
The summary is flagged as lossy so the agent verifies against the user's files instead of trusting it. No instruction that tells the agent to skip re-grounding after a compaction.

Impact: For a user who builds real business, legal, or software artifacts across many sessions, a compaction can silently erase confirmed work from the agent's memory, and the agent will then confidently act against work it has no record of. That is a data-integrity and trust problem, not a cosmetic one.

Environment: Product: Claude Cowork (desktop-app feature), desktop app version 1.22209.0.0 (Publisher: Anthropic PBC), paid Max plan
OS: Windows
Execution mode: local and cloud
Working folder: a connected OneDrive folder, long-running and multi-day sessions
Claude Code CLI: NOT installed
Underlying Claude Code runtime this session: 2.1.209; model: Claude Opus 4.8
Observed: the pattern has recurred since the end of June and has become increasingly worse in July and is becoming a daily recurrence.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Run a long, multi-session project in Cowork with a connected folder.
2. In an earlier session, have the agent build a small feature and confirm it, then move on to larger work.
3. Continue across sessions until one auto-compacts (long context).
4. Observe: the generated summary reads as complete but omits the earlier confirmed feature, and the injected prompt tells the agent to resume without re-grounding.
5. Ask the resumed agent to continue: it acts against the omitted feature because it has no record of it.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

1.22209.0.0

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Non-interactive/CI environment

Additional Information

Related (NOT filed here, part of a broader Cowork report I could not file as one on this repo)

Listed only as a reference so the full picture is on record. Each is its own issue:

1. Cloud sessions stage only a PARTIAL subset of the connected folder (about 28 of roughly 187 files) with no signal, so the agent wrongly reports files "missing" or "never saved."
2. The staged copy and the real device files diverge; bidirectional OneDrive sync causes version churn; no reliable "did my write reach the device" signal (data-loss risk on a naive write-back).
3. Cloud execution is ON by default (opt-out, not opt-in) for a feature that stages files to remote infrastructure.
4. Wide content (tables, code) is cut off on the right of the chat view with no horizontal scroll.
5. The composer grows and covers the chat history while typing a long message.
6. No agent-callable feedback hook that submits a user-authored report without sending the full conversation transcript.
7. The support intake rejected a valid report on a factual error ("Cowork appears to be a separate product or service"), so none of these reached the team.